[Go SDK] max_tokens not rewritten to max_completion_tokens for newer OpenAI models

[santoshkumarradha]

Summary

The Go SDK AI client always emits max_tokens in request bodies, but newer OpenAI models (o1-series, gpt-4.x-series) only accept max_completion_tokens and silently ignore or reject max_tokens, causing Go agents to lose their output-length cap.

Context

sdk/go/ai/request.go serializes the user-configured token limit as max_tokens. OpenAI's o1 and later 4.x model family dropped support for max_tokens in favor of max_completion_tokens. The Python SDK handles this via a per-model rewrite in litellm_adapters.py. Go has no equivalent rewrite pass: agents pointed at o1-mini, o1-preview, or gpt-4o variants silently send a parameter the model ignores, and the generated response may be longer or shorter than intended with no error returned. This is a silent correctness bug affecting any Go agent using newer OpenAI models.

Scope

In Scope

• Add a provider/model-aware rewrite pass in the Go SDK AI client that substitutes max_completion_tokens for max_tokens when the model requires it.
• The model list requiring the rewrite should be configurable or derived from a simple prefix/contains check (e.g. models starting with o1, o3, or matching OpenAI's documented list).
• The rewrite should only apply to OpenAI-compatible endpoints (detected via base URL or an explicit provider hint).

Out of Scope

• Rewriting other OpenAI-deprecated parameters — focus on max_tokens only.
• Supporting Anthropic or Gemini model quirks in this issue — OpenAI only.
• Changing the user-facing MaxTokens field name in RequestConfig — rewrite at serialization time, not at the API level.

Files

• sdk/go/ai/request.go — add model-aware serialization: emit max_completion_tokens instead of max_tokens for matching OpenAI models
• sdk/go/ai/client.go — apply the rewrite in the request-build path, using model name and base URL to detect when it's needed
• sdk/go/ai/request_test.go — tests: o1-mini request serializes max_completion_tokens; gpt-3.5-turbo request serializes max_tokens; non-OpenAI base URL uses max_tokens regardless of model name

Acceptance Criteria

• Requests to OpenAI o1-series and gpt-4.x models that require it include max_completion_tokens (not max_tokens) in the JSON body
• Requests to models that still use max_tokens are unaffected
• The model list / detection logic is documented in code comments and easy to extend
• Tests pass (go test ./sdk/go/...)
• Linting passes (make lint)

Notes for Contributors

Severity: MEDIUM

Reference sdk/python/agentfield/ai/litellm_adapters.py for the model list and rewrite logic used in Python. A simple strings.HasPrefix(model, "o1") or a small allow-list map[string]bool is sufficient — avoid pulling in a full model registry. If the OpenAI base URL detection is ambiguous, add an explicit Provider string field to ClientConfig so callers can declare intent.